On Tue, Nov 23 2021, Johannes Schindelin wrote:
> Hi Ævar,
>
> On Tue, 23 Nov 2021, Ævar Arnfjörð Bjarmason wrote:
>
>> [...]
>> $ make clean; make -C contrib/scalar test
>> [...]
>> CC hook.o
>> CC version.o
>> CC help.o
>> AR libgit.a
>> make[1]: Leaving directory '/home/avar/g/git'
>> SUBDIR ../..
>> make[1]: Entering directory '/home/avar/g/git'
>> * new link flags
>> CC contrib/scalar/scalar.o
>> LINK contrib/scalar/scalar
>> make[1]: Leaving directory '/home/avar/g/git'
>> make -C t
>> make[1]: Entering directory '/home/avar/g/git/contrib/scalar/t'
>> *** prove ***
>> error: GIT-BUILD-OPTIONS missing (has Git been built?).
>> t9099-scalar.sh .. Dubious, test returned 1 (wstat 256, 0x100)
>> No subtests run
>
> That's cute. You seem to have missed that this is `contrib/`? The
> assumption of pretty much _everything_ in there is that Git was already
> built.
>
> Try this at home: `make clean && make -C contrib/subtree/ test`
>
> Yep. It "fails" in the same way. "has Git been built?".
>
> So if that was all the evidence in favor of that misinformation "Scalar's
> build is broken! Broken, broken, BROKEN!", I think we can now let it rest.
> At last.
No, it doesn't fail in the same way. Really, it seems like you're either
not fully reading through E-Mails before replying, or entirely
misunderstanding what I'm saying. I highlighted the difference in
"[...]that's not broken because[...]" below the context you're quoting.
I'm specifically pointing out the difference between how these act:
make clean; make -C contrib/subtree/ test
make clean; make -C t
Which fail right away without trying to build anything, and how your:
make clean; make -C contrib/scalar test
Won't fail right away, but get most of the way towards building what it
needs at the top-level.
So yes, I fully agree with your contrib/subtree example, but it's making
my argument for me.
As I pointed out in the just-sent [1] the main issue is that your latest
iteration added "make -C contrib/subtree/ test" in the wrong place, and
we thus ended up running those tests in places we didn't intend. The
main conflict is that semantic conflict.
But that failure also highlighted that contrib/scalar/Makefile will run
a build of the top-level C code, only to error out with
"GIT-BUILD-OPTIONS".
That specifically is what I consider broken. It should either actually
work and properly build its dependency, or not even try.
Either one would be fine in this context, it's the in-between that's
clearly (to me at least) broken. Working software should either attempt
a task and succeed, or not attempt the task at all.
1. https://lore.kernel.org/git/211123.86ilwjujmd.gmgdl@xxxxxxxxxxxxxxxxxxx/
